What changed
AWS published a LangSmith-based pattern for tracing and evaluating long, tool-using agents across trajectories, intermediate decisions, outputs and repeated datasets.
Why it matters
Deep agents can reach the right answer through unsafe or wasteful paths, so outcome scores need tool, trajectory, latency and cost evidence.
What to do
Trace one multi-tool workflow, label forbidden and unnecessary calls and compare outcome, path quality, latency and cost across three repeated runs.
Caveat
The example combines vendor services and judge-based evaluation, which can add lock-in, data-governance work and grader instability.